The Match Tracking Anomaly and its Minimisation in the Fuzzy ARTMAP Neural Network


Author: Guszti Bartfai
Source: GZipped PostScript (91kb); Adobe PDF (375kb)

This paper looks at the anomalous situation that can arise in the ARTMAP and Fuzzy ARTMAP neural networks --- as a result of their ``match tracking'' process --- whereby a current input--target association cannot be learned because the network is unable to find an alternative category for an input pattern that matches its selected category prototype perfectly. We show that this Match Tracking Anomaly (MTA) cannot be avoided in the Fuzzy ARTMAP network if its ARTb vigilance level is less than 1. To alleviate this problem, we propose two improvements to the Fuzzy ARTMAP learning algorithm. One of them is concerned with the timing according to which input patterns and corresponding target outputs are processed by the network. The other one is the explicit overwriting of an existing association between an input and an output category in case the input is matched perfectly and yet the network's prediction is wrong. Both of these modifications are needed to reduce the occurrence of MTA during learning, and eliminate it altogether in a trained network (on a finite training set). As a result, training time is also reduced, which is demonstrated through the performance of the network on a machine learning benchmark database.

[Up to Computer Science Technical Report Archive: Home Page]